Ryan Huang retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me.
I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA.
I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation.
Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO.
Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours.
Jensen
Ryan Huang retweeted
Last Oct in Seoul, SIGOPS held a strategic workshop on systems research,which I chaired and co-organized. Several generations of systems community joined. We debated AI+systems, reliability, cloud/AI systems, and the OS agenda. Report: sigops.org/2026/the-seoul-re… Feedback welcome.
1/ AI coding agents are getting remarkably good at writing code. But can they troubleshoot a failing distributed system and find the root cause in its source code?
We built DistDebug-Bench to find out (NeurIPS 2026) 🧵
5/ 🔦 Tunnel vision: agents lock onto their first plausible explanation and rarely recover.
Paper, findings, and leaderboard: distdebug.com/
1
94
6/ The benchmark, evaluator, agent traces, and leaderboard are all public. We'd love to see your model/ agent on the board!
⭐ github.com/OrderLab/distdebu…
w/ @_MattJiang_ Chenglin Liu, John Rose, Mason Miller, Ruthesh Thavamani, Juntao Wu, Yinwei Huang, and Jiahao Xu
1
128
🎉 Excited that our work, OpGuard, was nominated for OSDI26 best paper award!
Debugging is notoriously hard. In LLM training, it is even harder: a tiny error may remain invisible for a long time, surfacing only after it has propagated across GPUs, operations, and model state.
How can we quickly determine when and where things first went wrong? OpGuard introduces bitwise alignment as a precise debugging primitive that pinpoints the first divergence between two training executions. In production, it has reduced debugging time from days to minutes.
2
1
223
My student Ziming will present the work tomorrow, July 14, in the "Training Reliability and Silent Errors" session, 2:00–3:20 PM. Please stop by to hear Ziming present the work!
usenix.org/conference/osdi26…
1
162
Ryan Huang retweeted
We’re @UMich researchers looking for 20-minute conversations with people who build, train, fine-tune, evaluate, or operate ML systems.
Despite major progress in ML infrastructure reliability, many practitioners still point to the correctness of training and evaluation as a central bottleneck.
We’re trying to understand what gaps remain, how teams detect and handle them today, and where better tooling could help.
If this touches your work, please reply or DM. Referrals are also appreciated.
#MLSys #MachineLearning #AIInfrastructure #MLOps
Ryan Huang retweeted
Overload control is usually built around a bad assumption. Most systems watch global signals like queue length or tail latency and react at the front door by throttling new arrivals or dropping random requests. This works when CPU or network is the bottleneck. It fails when the real problem sits deeper inside the application. In practice, many overload incidents come from one or two requests that monopolize an internal logical resource like a buffer pool, a lock, or a thread-pool queue. These rogue whales distort the system. One ill-timed dump query can thrash the buffer pool and cut throughput in half. One backup thread plus a heavy table scan can stall writes in MySQL. None of this shows up in CPU metrics.
The Atropos paper (SOSP’25) offers a simple fix. Instead of punishing victims at admission time, Atropos watches how tasks actually use these internal resources and cancels the ones causing the collapse.
The key observation is that most real systems already have safe cancellation hooks. What’s been missing is a principled way to decide whom to cut. Atropos supplies that missing piece and shows that killing a single rogue whale can unblock the whole system.
Ryan Huang retweeted
After six wonderful years at UIUC @siebelschool, I’m thrilled to announce that I will join the University of Toronto ECE @eceuoft as an assistant professor in July 2026. I’m actively looking for PhD students in fall 2026. Drop me an email if you are interested in working with me.
Ryan Huang retweeted
🔍 Meet TrainCheck, an open-source tool from #UMich CSE led by Prof. @ryanphuang that catches silent errors in deep learning—bugs that quietly wreck model quality. It found 18/20 real-world errors in 1 iteration + 6 unknown bugs.
▶️cse.engin.umich.edu/stories/…
#TrainCheck
Ryan Huang retweeted
Congratulations to the authors of Basilisk: Using Provenance Invariants to Automate Proofs of Undecidable Protocols on receiving a ⭐ Best Paper Award ⭐ at OSDI 2025! 👏
🔗 Read more: cse.engin.umich.edu/stories/…
#OSDI2025 #SystemsResearch #BestPaper #CSEMichigan
Ryan Huang retweeted
🚨 New resource for ML systems folks!
We release a curated reading list on ML reliability, including silent errors, testing, fault tolerance, and more.
If you’re building more robust, debuggable ML systems, this list is for you 👇
🔗 github.com/OrderLab/awesome-…
PRs are welcome!
🚨 Excited to share that our group will present two papers at OSDI '25 next week!
🎯 TrainCheck: Automatically catches silent errors during deep learning training by inferring and enforcing training invariants.
* arxiv.org/abs/2506.14813v1, [Code]( github.com/OrderLab/TrainChe…)
1
16
761
🔍 T2C: Turns system tests into semantic checkers to detect failures in distributed systems.
* [Paper](web.eecs.umich.edu/~ryanph/p…), [Code](github.com/OrderLab/T2C)
#DistributedSystems
2
202
Ryan Huang retweeted
Come join the Artifact Evaluation Committee for #osdi25 & @usenixtatc25, and help promote reproducibility in our Systems community.
Ryan (@ryanphuang) and Tianyin (@tianyin_xu) will make this process fun, smooth, and lightweight (1-2 artifacts per reviewer).
We are seeking members to join the Artifact Evaluation Committee for #OSDI25 & #USENIXATC25.
Help promote reproducibility, and engage with cutting-edge systems research!
Please apply by April 17th via forms.gle/187NkvkWyNtBHDsu5.
#SystemsResearch #Reproducibility #OSDI #ATC









